← Back to selected work
Case Study 06 / Human-Centred AI Exploration · Watchathon

Watching together. Staying in control.

Watchathon explores how friends and family could share a movie night across distance through synchronised viewing, individual language choices and optional AI-assisted social cues. What began as a solo Goldsmiths UX-in-AI academic concept has been continued as an end-to-end, clearly labelled portfolio concept prototype.

Explore the interactive concept prototype Fictional demo · no streaming, camera, microphone or AI

Original 2026 illustration of the concept: one shared screen, three fictional viewers with their own language settings and manually chosen reactions. No real film, people or interface.
Project
Watchathon · AI-assisted remote co-watching
Academic context
Goldsmiths · MSc UX Engineering · UX in AI
My role
Solo academic project · Research and proposal
Current status
Original concept documented · Portfolio concept prototype created (2026) · Not shipped

Connection is human. AI should know its limits.

The original idea proposed recognising viewers' emotions through a camera and microphone. Revisiting it through a contemporary human-centred AI lens raises a better question: could technology support shared presence while letting people decide what they reveal about themselves?

01 / Original research & proposalHistorical academic project · Documented, not reconstructed results

A movie night, across the miles.

The original Watchathon proposal was concerned with something simple: watching a film with the people you care about when you live in different places. It imagined shared playback, different language choices for each participant and camera/microphone communication to bring reactions into the experience.

The human problem

How might people share the feeling of watching together while physically apart?

Academic problem framing, not a measured population-level unmet need.
The original AI idea

Use camera and microphone input to estimate emotional reactions and share cues with other viewers.

A speculative proposal in the original project. No trained emotion model, prediction accuracy, or live implementation is evidenced.

Research / Questionnaire

25 responses

I conducted and analysed a questionnaire with 25 participants during the original solo academic project. The raw response data is no longer available, so I cannot responsibly publish response breakdowns, charts or exact percentages.

Raw data unavailable
Research / Interview

Qualitative exploration

The original project write-up describes interview work alongside demographic research and the questionnaire. Original notes, question wording and recruitment details are not currently available; no quotations or detailed interview findings are reproduced.

Notes unavailable · no quotes
Research / Reported theme

Language choice

The original write-up reports that most questionnaire respondents preferred watching in their own language. The underlying count is missing, so this remains an attributed qualitative account rather than a new, independently checked statistic.

Count missing · attributed account

The evidence can support

  • An owner-confirmed sample size: 25 questionnaire responses
  • Themes attributed to the original write-up
  • The problem framing and the original proposal

It cannot support

  • Percentages, charts or response breakdowns
  • Participant quotes or personas
  • Claims about demand, behaviour or impact

Source: six owner-supplied screenshots of the historical Watchathon project presentation, including the problem, process, proposal, Emotional AI limitations and systems diagram; owner confirmation that this was a solo academic project and that 25 questionnaire responses were collected. The screenshots are private source material and are not republished here.

02 / Original systems thinkingAcademic proposal · Not a technical build

One shared session. Several moving parts.

The original systems diagram placed a host and remote viewers around a common Watchathon interface. Connectivity, camera, microphone and an Emotional AI layer were all considered parts of the experience. This was a conceptual architecture that helped explain the idea, not a completed infrastructure or implemented AI service.

01 / People

Host and viewers

One person organises the watch session while invited friends or family join remotely.

02 / Shared viewing

Playback and language

A shared watch context with individual language preferences, subject to actual content rights and availability of separate audio or subtitle tracks.

03 / Proposed AI

Presence cues

The initial idea used camera/microphone signals to infer and communicate emotional state. That capability was proposed rather than built or validated.

PersonHostCreates the session and invites peopleCameraMicrophone
SharedWatchathon interfaceOne shared watch contextConnectivity
PersonRemote viewerOwn language preferenceCameraMicrophone
PersonRemote viewerOwn language preferenceCameraMicrophone
Proposed layer · not built or validatedEmotional AIWould have interpreted camera and microphone signals as emotional reactions
Academic conceptual model · not deployed architecture A 2026 portfolio redrawing of the functions in the original systems diagram. Not a copy of the original artwork and not a technical build.

Originality boundary: these are the functions described in the historical presentation, not a claim of an existing streaming integration, live video platform, working emotion classifier or completed product.

03 / Human-centred AI critiqueRetrospective design perspective · Not a newly completed study

Can a camera tell you how someone feels?

A smile, silence or a turned-away face does not reliably reveal a person's inner emotional state. The original academic work already raised cultural, geographic and individual differences as challenges for emotion recognition. The portfolio continuation treats these limits as design constraints rather than positioning Emotional AI as an accurate way to read people.

What the product might notice

Observable cues, uncertainty and voluntarily shared reactions.

Hypothetical future capabilities; no sensor model or classifier is implemented.
What AI must not assert

“I know exactly how you feel.”

Facial movements vary by context and person. The design should never present an emotion estimate as a fact, diagnosis or measure of wellbeing.

Observed: Corners of the mouth lift

Model's guess: Amused?

Might actually feel: A polite smile while worrying about tomorrow

Observed: Eyes turn away from the screen

Model's guess: Bored?

Might actually feel: Listening closely to the dialogue

Observed: Quiet for a long stretch

Model's guess: Sad?

Might actually feel: Tired after a long day — and happy to be there

A signal is not a feeling. Observable cues can be measured; a model's interpretation of them is a guess; what a person feels is theirs to share. Illustrative examples · not data from the original study

The key design reframing is to ask whether sharing a reaction is valuable because a person chooses to share it, rather than because a system claims to detect and announce it. Manual reactions become the reliable starting point; speculative AI assistance remains optional, reversible and subject to further research.

External scientific context, added for the portfolio continuation: Barrett and colleagues (2019), Emotional Expressions Reconsidered. This publication is background evidence about emotion inference; it is not an original Watchathon study result.

04 / Product principlesNew portfolio continuation · Proposed decisions

AI should participate on people's terms.

Manual-first01 / Human connection

People before predictions

Make it easy to talk, react and watch together even when AI features are turned off.

Consent02 / Genuine choice

Off by default

Any future affective analysis would need separate, informed, granular opt-in. Participation in a session must not require it.

Uncertainty03 / Uncertainty

Never pretend to know

Do not interpret ambiguous faces or voices as a definite emotional state. Allow uncertainty, correction and no result.

Human agency04 / Agency

Let people choose

Use explicit sharing controls so a participant can select what reaction, if any, other viewers see.

Accessibility05 / Inclusion

Many ways to be present

Support text-only participation, captions, accessible controls, variable expressions and different language needs.

Privacy & transparency06 / Plain language

Explain the feature

State what the system proposes to process, what it cannot know, when data would be shared and how to stop it.

05 / Product definitionNew portfolio continuation · Proposed scope

A watch room with room for real people.

The concept continuation turns Watchathon into a defined experience: an organiser sets up a session, friends or family join, each viewer chooses available language options, and everyone can react through intentional, controllable interactions. The AI layer is explored as an optional addition, not the prerequisite for watching together.

Illustrative job to be done

When we live far apart, help us share a film night without losing the feeling of being together.

A constructed JTBD based on the concept. Not a quotation or verified finding from the 25 participants.
Concept feature set

Room creation, joining and invitations; per-viewer language preferences; shared viewing states; manual reactions and optional presence; controllable AI concept cues.

These are product requirements for a new concept prototype, not evidence of features already built or integrations with streaming providers.

Illustrative scenario · not a participant story

Saturday evening. Three friends in different cities pick a film. One prefers Hindi subtitles, another wants the original audio, and nobody wants to keep a camera on all night. They still want to laugh at the same moments.

  1. Pick a film and a room
  2. Each chooses a language
  3. Decide what to share
  4. Watch and react

Core · manual-first

  • Create a room and join with a code
  • Per-viewer language choices
  • Shared watch state
  • Manual reactions
  • Leave or end a session

Optional · opt-in only

  • Scripted AI concept cue, with explanation, correction, dismissal and an off switch

Out of scope for the concept

  • Synchronised streaming
  • Licensed films
  • Voice or video calling
  • Camera or microphone processing
  • Emotion inference
  • Accounts

The portfolio demonstration uses fictional film content and simulated room/session states. Actual synchronised playback, rights-managed movies, independent audio tracks, account management, video calling and emotion processing remain outside its functional scope.

06 / End-to-end experienceProposed flow · Not observed participant behaviour

From invitation to shared memory.

The experience is designed as a complete but fictional journey. It lets a visitor move through a coherent prototype rather than only inspect a set of static screens.

01 / Plan

Create a room

Pick a fictional film, name the room, set the host and prepare an invite.

02 / Invite

Join together

Share an illustrative room code or use a demo invite; reveal who has joined using fictional people only.

03 / Personalise

Choose a language

Select from sample audio/subtitle options, clearly noting that real track availability depends on content licences.

04 / Consent

Decide what to share

Review camera/microphone presence separately from any speculative emotion analysis and choose a no-AI path.

05 / Co-watch

React your way

Experience a simulated watch view, pause/resume state, manual reactions and an optional mock AI-cue explanation.

06 / Reflect

End and evaluate

Finish the demo with a sample post-watch view and optional feedback questions, not a claimed social-impact result.

AI and privacy choices along the way

  1. 01No account, camera or microphone needed
  2. 02Guests are fictional; no invitation is sent
  3. 03Each viewer chooses their own language
  4. 04No AI cues defaultTry the scripted AI concept needs acknowledgement
  5. 05React manually · hide others' reactions · turn AI off at any time
  6. 06Session choices aren't saved · restart from scratch

All stages in this proposed journey are created for the new portfolio concept. They have not been validated as the actual behaviour or preferences of original research participants.

07 / AI experience modelHypothetical design logic · No trained model

Observe less. Ask more.

The original proposal suggested emotional recognition through camera and audio cues. The continuation distinguishes signals the user explicitly supplies from signals an AI might attempt to interpret. It then designs guardrails around any future inference rather than assuming a model can reliably distinguish feelings.

Input / Available in concept demo

Direct reactions

User selects “Laughing”, “Surprised”, “Quiet moment” or no reaction. These are explicit, user-provided signals and require no emotion-recognition AI.

Input / Speculative

Sensor-derived cues

Camera and microphone interpretation was proposed in the original research. It is not implemented in the portfolio prototype and would need separate consent, feasibility and scientific evaluation.

Output / Guardrails

Optional suggestions

If a future system offered a cue, it would need uncertainty handling, a user-visible explanation, an easy dismiss/correct control and an explicit decision about sharing.

Illustrative decision boundary

No consent? No analysis. Uncertain? No inference shown.

A desired product policy, not a guarantee supplied by an existing AI service.
What the prototype simulates

Opting in or out, seeing why a cue is uncertain, correcting a suggestion and deciding whether to share.

UI states use prewritten fictional examples. The prototype must never suggest it is reading the visitor's face, microphone or emotions.

Manual path · works without AI

  1. Input · availableYou tap a reaction
  2. ResultShared because you chose it

AI concept path · hypothetical

  1. Input · speculativeCamera or microphone cue Not implemented
  2. GateConsent given?No → no analysis
  3. Processing · hypotheticalA model makes a guess No model exists
  4. GateToo uncertain?Yes → nothing shown
  5. OutputA suggestion, never a fact
  6. The person decidesShare · correct · dismiss · turn off
Conceptual interaction model · no trained model, dataset, accuracy or inference The prototype only simulates the final two steps, with one prewritten example.

External risk-management reference for the portfolio continuation: NIST AI Risk Management Framework (validity, reliability, transparency, privacy and fairness). This is external guidance, not a certification or an original Watchathon result.

08 / Privacy, ethics & inclusionProduct requirements to validate · Not shipped safeguards

People should control their own presence.

Emotion cues can be intimate, misleading or socially uncomfortable. A person might be tired rather than sad, look away while listening, or communicate differently from what a model expects. Cultural variation, disability, neurodivergence and situational context are reasons to challenge simplistic interpretations.

Consent

Specific and reversible

Separate camera, microphone, hypothetical analysis and sharing choices. A room should remain usable without AI features.

Transparency

Tell people what's happening

Explain whether a cue is manually shared or a hypothetical AI suggestion, and label fictional demo content unambiguously.

Privacy

No false storage promises

The portfolio demo does not collect real camera, microphone, login or emotion data. A real implementation would require an independently verified data-retention and processing design.

Fairness

Test across differences

Do not treat expressions as universal. Any future AI experiment would need explicit opt-in studies, subgroup analysis and expert oversight before release consideration.

Accessibility

Always offer alternatives

Use text, captions, keyboard access, readable status and manual reactions; camera-off and quiet participation must work as first-class paths.

Social safety

Respect boundaries

Let viewers hide cues, mute others, leave a room and report an uncomfortable interaction; provide visible state and confirmation.

Consent anatomy, taken from the 2026 concept prototype.
  1. 1Hypothetical use, stated plainlyWhat a real product would process, and why that is unreliable.
  2. 2What this demo actually doesOne prewritten example; no sensing.
  3. 3Who sees itPrivate by default; sharing is a separate decision.
  4. 4How to stopRevocable at any time, from the room itself.
  5. 5Off by defaultThe no-AI path is pre-selected and complete.
  6. 6Equal-weight choicesNo nudging towards opting in.
  7. 7Explicit acknowledgementOpting in needs a deliberate confirmation.

Camera-off path

Every step works without a camera or microphone.

Uncertainty messaging

Cues say "might" and explain what they cannot know.

Correction

Choose a different reaction, or keep it private.

Revocation

Turn AI cues off mid-session; nothing is recorded.

Inclusive alternatives

Manual reactions, captions, language choice and keyboard access.

Release risks to evaluate before any launch Not assessed results

  • Misreading people's feelings
  • Social pressure to opt in
  • Uneven performance across faces, voices and cultures
  • Feeling watched in a social space
  • Excluding people who express themselves differently
  • Misuse or retention of sensitive data
09 / Wireframes, UI & interactive prototype2026 portfolio continuation · Created and tested in the browser

Make the reasoning feel usable.

The original submission was a research proposal with a system diagram, not a documented clickable interface. The 2026 continuation created wireframes, high-fidelity screens and a genuinely navigable, browser-based demo. They are explicitly dated as a new portfolio continuation, never misrepresented as coursework artefacts or a released Watchathon product.

Low fidelity / Created

Structure before styling

Document room creation, invitations, language setup, AI/privacy choices, the watch-room shell and session wrap-up as connected wireframes.

High fidelity / Created

A calm watch interface

Develop original screens for onboarding, room setup, lobby, consent, active viewing, manual reactions, AI explanation/settings and post-session reflection.

Interactive / Created

Real browser controls

Build a static, offline-friendly concept demo with working steps, state, optional privacy toggles, manual reactions and a simulated session. No working media stream, web API or emotion-recognition model.

2026 portfolio continuation · Interactive

Explore the interactive concept prototype

Create a room, invite fictional guests, choose languages, decide about AI and react in a simulated session. The no-AI route is complete on its own.

Explore the interactive concept prototype

Truly working in the browser

  • Step-by-step navigation with Back and Restart
  • Room name and film choice, with validation, carried through every screen
  • Language, subtitle and size choices with a live preview
  • Consent choice, off by default, with acknowledgement
  • Play and pause timer, manual reactions, hiding others' reactions
  • Correcting, dismissing and turning off the scripted cue
  • Session summary and optional feedback prompts

Simulated or fictional

  • Films, guests and room codes
  • Playback: no video is streamed or synchronised
  • Guest joining and guest reactions
  • The AI cue: one prewritten example, no camera, microphone, model or inference
  • Feedback: shown only on screen, never stored or sent

Wireframes

2026 portfolio continuation · low fidelity

Structure before styling: each wireframe fixes the order of decisions, especially where AI and privacy choices sit.

WF1Room setup
  • Room name field with help text
  • Film choice as large cards
  • Back and Continue always paired
WF2Invite
  • Demo code, clearly labelled
  • Who is in the room
  • Join is simulated, never sent
WF3Languages
  • Audio, subtitles and size as separate groups
  • Live preview of the choice
  • Licensing note beside it
WF4Consent
  • Manual reactions separated from AI
  • AI off by default; equal-weight options
  • Acknowledgement only if opting in
WF5Watch room
  • Shared screen first, room panel beside it
  • Reactions one tap away
  • AI panel collapsible and switchable off
WF6End of session
  • Plain summary of what happened
  • Optional, skippable feedback
  • Restart clears everything

High-fidelity screens

2026 portfolio continuation · fictional content · from the working prototype

The polished screens use the same components as the interactive prototype, so what you see here is what you can click there.

UI1Welcome
UI2Room setup
UI3Language selection
UI4Consent
UI5Lobby
UI6Watch room
UI7AI cue explanation
UI8End of session

Prototype: a separate accessible page at work/watchathon-prototype.html. It distinguishes functional browser interactions from simulated product capabilities, as listed above.

10 / Evaluation strategyProposed study, no new participants tested

Measure clarity before excitement.

A prototype creates an opportunity to test decisions, not proof that they work. The next research stage should examine whether people understand the AI disclosures, can complete the watch setup without assistance and can stop or correct cues without being pressured into sharing.

Task / Setup

Can people join?

Proposed measure: completion of room setup and language preferences, including the no-camera/no-AI path.

Not yet run
Task / Consent

Do people understand?

Proposed measure: correctly identifying which data a feature would use, who could see a cue and how to turn it off.

Not yet run
Task / Recovery

Can people correct it?

Proposed measure: ability to dismiss an unclear suggestion, change consent and recover from unavailable languages or playback desynchronisation.

Not yet run

These are research hypotheses and proposed measures. No new participant study, prototype task success rate, model accuracy, satisfaction score, retention number or product impact has been collected for this portfolio continuation.

11 / MVP to release-readinessForward-looking shipping strategy · Not a launched product

Know what to ship. Know what not to.

A responsible product cycle includes discovery, a clickable prototype, feasibility testing, an MVP, pre-release validation and release safeguards. This project will use that sequence to demonstrate product judgement, rather than claim all those stages happened in the original academic submission.

Completed · historicalPhase 00 / Completed historically

Research and concept

Solo academic framing, the 25-response questionnaire, interview work described in the historical write-up, an AI-assisted co-watching proposal and system diagram.

Created · 2026 portfolioPhase 01 / Portfolio continuation

Prototype and critique

Wireframes, original screens and an interactive static demonstration of consent, reactions and session states. All simulated functionality is disclosed.

Future gate · not startedPhase 02 / Future MVP gate

Validate basic co-watching

First test room setup, invitations, manual reactions, language availability and rights-compliant content. No emotion-recognition AI required for the MVP.

Future gate · not startedPhase 03 / Future risk gate

Decide whether AI belongs

Only consider sensor-based AI after scientific feasibility, privacy review, informed-consent testing, accessibility assessment, bias evaluation and evidence of incremental user value.

Future gate · not startedPhase 04 / Future launch gate

Production readiness

Content/streaming licences, security and data governance, performance, accessible interactions, moderation/abuse response, incident handling and operational support.

Proposed success measures for a future product: invite-to-join completion, setup completion, consent comprehension, feature disable/correction, accessible task success, session completion and repeat-use interest. All are proposed metrics with no measured baseline or results.

12 / ReflectionOriginal ambition · Contemporary critical perspective

The product should make room for uncertainty.

My original idea set out to make distance feel smaller through shared films, language choice and Emotional AI. Revisiting it, the central question has shifted. The most interesting design challenge is not whether a machine can label every facial expression, but whether a system can support meaningful, voluntary interaction without presuming to know what people feel.

This 2026 design stage makes that thinking tangible: a manually controlled watch experience first, an optional simulated AI layer, explicit consent and a proposed validation and release strategy. It is a demonstration of end-to-end product thinking — not evidence of an implemented model, licensed streaming service or production launch.

13 / Evidence and authorshipA clear boundary between original and new work

What's original, what's new, and what's proposed.

Original academic work

Documented

Solo UX-in-AI research and concept, questionnaire with 25 participants, qualitative interview work described in original screenshots, the shared-viewing/language/emotion-AI proposal and a system diagram.

Portfolio continuation

Created in 2026

Human-centred AI critique, product requirements, flows, wireframes, new screen designs, interactive simulated prototype and test/release plan. These are later work and are labelled as such.

Not claimed

No invented outcomes

No accessible raw questionnaire dataset, exact percentages, live viewing integration, trained model, tested emotion accuracy, production deployment, measured happiness or claimed subscription demand.

Original source: the project owner's six historical Watchathon screenshots and direct confirmation of the solo project and 25 responses. The original screenshots include third-party media/illustrations and are not copied. External references added in this later continuation: Barrett et al. (2019) on limits of emotion inference, and the NIST AI Risk Management Framework. These external sources are not claimed as findings from the original academic study.